This is a FRA funded project where the main goal is to design a system to predict flaws and cracks in the rail track and also its remaining life. The proposed solution is computer-vision based and performs image-anomaly localization on the B-scan data. This is an ongoing project.
This is an ongoing ARM funded project where the goal is to automate finishing tasks in manufacturing processes. In this project, we consider th specific case of edge deburring. My contribution is in developing the algorithm to perform edge detection on the mesh and then generate a toolpath along which the robot will perform the deburring operation.
This is a DARPA Physics of AI project where the broad goal is to integrate physics constraints into neural networks so as to learn representations that respect these constriants. One of the applications we were addressing was to perform 3D recosntruction of molecule given various 2D projections of the same. The imaging procedure, Cryo-EM is very noisy and is hard to obtain many image. There is also no control over the angle at which the molecule is images as they are taken randomly. Furthermore, the molecule considered also has icosohedral symmetry. The goal of the project is to first estimate the 3D (euler or quaternion) angles of each of these images, and then perform 3D reconstruction of the molecule. A maximum likelihood based approach (Hetero) is first used to estimate these angles which serves as a ground truth and then a neural network is learned to regress these angles. We also develop a disentangled VAE to learn the angles explicitly at the latent space and then use latent space interpolations to generate images of unseen angles and in-turn obtain better 3D reconstruction.
The main goal is to create an automated tracking and traceability solution for part accountability in manufacturing processes. Infrared imaging was used to scan the objects and we developed a YoLov3 based object detection pipeline for the same.